[PATCH] dcache: ensure d_flags & d_inode are consistent in lookup_fast()

[Hou Tao <houtao1@huawei.com>]

After extending the size of dentry from 192-bytes to 208-bytes
under aarch64, we got oops during the running of xfstests generic/429:

  Unable to handle kernel NULL pointer dereference at virtual address 0000000000000002
  CPU: 3 PID: 2725 Comm: t_encrypted_d_r Tainted: G      D           5.1.0-rc4
  pc : inode_permission+0x28/0x160
  lr : link_path_walk.part.11+0x27c/0x528
  ......
  Call trace:
   inode_permission+0x28/0x160
   link_path_walk.part.11+0x27c/0x528
   path_lookupat+0x64/0x208
   filename_lookup+0xa0/0x178
   user_path_at_empty+0x58/0x70
   vfs_statx+0x94/0x118
   __se_sys_newfstatat+0x58/0x98
   __arm64_sys_newfstatat+0x24/0x30
   el0_svc_common+0x7c/0x148
   el0_svc_handler+0x38/0x88
   el0_svc+0x8/0xc

If we revert the size extension of dentry, the oops will be gone.
However if we just move the d_inode field from the begin of dentry
struct to the end of dentry struct (poorly simulate the way how
__randomize_layout works), the oops will reoccur.

The following scenario illustrates the problem:

precondition:
* dentry A has just been unlinked and becomes a negative dentry
* dentry A is encrypted, so it has d_revalidate hook: fscrypt_d_revalidate()
* lookup process is looking A/file, and creation process is creating A

lookup process:                 creation process:

lookup_fast
    __d_lookup_rcu returns dentry A

    d_revalidate returns -ECHILD

    d_revalidate again succeed
				__d_set_inode_and_type
				dentry->d_inode = inode
				WRITE_ONCE(dentry->d_flags, flags)

    d_is_negative(dentry) return false
    follow_managed doesn't nothing
    // inconsistent with d_flags
    d_backing_inode() return NULL
    nd->inode = NULL

may_lookup()
    // oops occurs
    inode_permission(nd->inode

The root cause is the inconsistency between d_flags & d_inode
during the REF-walk in lookup_fast(): d_is_negative(dentry)
returns false, but d_backing_inode() still returns a NULL pointer.

The RCU-walk path in lookup_fast() uses d_seq to ensure d_flags & d_inode
are consistent, and lookup_slow() use inode lock to ensure that, so only
the REF-walk path in lookup_fast() is problematic.

Fixing it by adding a paired smp_rmb/smp_wmb between the reading/writing
of d_inode & d_flags to ensure the consistency.

Signed-off-by: Hou Tao <houtao1@huawei.com>
---
 fs/dcache.c | 2 ++
 fs/namei.c  | 7 +++++++
 2 files changed, 9 insertions(+)

diff --git a/fs/dcache.c b/fs/dcache.c
index aac41adf4743..1eb85f9fcb0f 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -316,6 +316,8 @@ static inline void __d_set_inode_and_type(struct dentry *dentry,
 	unsigned flags;
 
 	dentry->d_inode = inode;
+	/* paired with smp_rmb() in lookup_fast() */
+	smp_wmb();
 	flags = READ_ONCE(dentry->d_flags);
 	flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
 	flags |= type_flags;
diff --git a/fs/namei.c b/fs/namei.c
index dede0147b3f6..833f760c70b2 100644
--- a/fs/namei.c
+++ b/fs/namei.c
@@ -1628,6 +1628,13 @@ static int lookup_fast(struct nameidata *nd,
 		return -ENOENT;
 	}
 
+	/*
+	 * Paired with smp_wmb() in __d_set_inode_and_type() to ensure
+	 * d_backing_inode is not NULL after the checking of d_flags
+	 * in d_is_negative() completes.
+	 */
+	smp_rmb();
+
 	path->mnt = mnt;
 	path->dentry = dentry;
 	err = follow_managed(path, nd);
-- 
2.16.2.dirty